[PATCH] ntfs: fix sector conversion on 4Kn devices

[PATCH] ntfs: fix sector conversion on 4Kn devices

[朱 天浩]

The NTFS driver was converting byte offsets to bio sector numbers using
the filesystem block size (sb->s_blocksize_bits).  However, block layer
sector numbers - bio->bi_iter.bi_sector and sector_t - are always in
units of 512 bytes, independent of the filesystem block size.

On a device with a 4096 byte logical block size (4Kn), sb_min_blocksize()
lifts s_blocksize to 4096, so every MFT writeback bio built by
ntfs_sync_mft_mirror(), write_mft_record_nolock() and
ntfs_mft_writepages() ended up with a sector number eight times too
small.  The resulting LBA is not only wrong, it is usually not a
multiple of the logical block size either, and some controllers reject
the write outright, e.g.:

  ntfs: (device nvme0n1p3): ntfs_sync_mft_mirror(): I/O error while
  writing mft mirror record 0x3!
  ntfs: (device nvme0n1p3): write_mft_record_nolock(): I/O error while
  writing mft record 0x3!  Marking base inode as bad.

Reproduced on an Apple SSD AP0256J (4Kn) with a volume using 4096 byte
sectors, 4096 byte clusters and clusters_per_mft_record = 1; after the
fix a 13 MB file written to the volume is byte-for-byte identical
after a umount/mount cycle.

Fixes: 40796051991d ("ntfs: update in-memory, on-disk structures and headers")
Signed-off-by: Zhu Tianhao <zhutianhao75@hotmail.com>
---
 fs/ntfs/ntfs.h | 15 ++++++++++++---
 1 file changed, 12 insertions(+), 3 deletions(-)

diff --git a/fs/ntfs/ntfs.h b/fs/ntfs/ntfs.h
index df5a75d506f6..b66ce62f67c9 100644
--- a/fs/ntfs/ntfs.h
+++ b/fs/ntfs/ntfs.h
@@ -20,6 +20,7 @@
 #include <linux/smp.h>
 #include <linux/pagemap.h>
 #include <linux/uidgid.h>
+#include <linux/blkdev.h>      /* For SECTOR_SHIFT. */
 
 #include "volume.h"
 #include "layout.h"
@@ -71,7 +72,11 @@
 #define NTFS_CLU_TO_POFS(vol, clu) (((u64)(clu) << (vol)->cluster_size_bits) & \
 				    ~PAGE_MASK)
 
-#define NTFS_B_TO_SECTOR(vol, b) ((b) >> ((vol)->sb)->s_blocksize_bits)
+/*
+ * bio->bi_iter.bi_sector and sector_t are always in 512-byte sectors, even
+ * when the filesystem block size is larger (for example 4KiB).
+ */
+#define NTFS_B_TO_SECTOR(vol, b) ntfs_bytes_to_sector(vol, b)
 
 enum {
 	NTFS_BLOCK_SIZE		= 512,
@@ -154,11 +159,15 @@ static inline u64 ntfs_cluster_to_poff(const struct ntfs_volume *vol,
 	return (clu << vol->cluster_size_bits) & ~PAGE_MASK;
 }
 
-/* Convert byte offset to sector (block) number. */
+/* Convert byte offset to a block layer sector number.  Those are always in
+ * units of 512 bytes, never in units of sb->s_blocksize, which is raised to
+ * the device logical block size on 4Kn devices.
+ */
 static inline sector_t ntfs_bytes_to_sector(const struct ntfs_volume *vol,
 		u64 bytes)
 {
-	return bytes >> vol->sb->s_blocksize_bits;
+	(void)vol;
+	return bytes >> SECTOR_SHIFT;
 }
 
 /* Global variables. */
-- 
2.51.0

Re: [PATCH] ntfs: fix sector conversion on 4Kn devices

[liubaolin]

在 2026/9/3 20:44, 朱 天浩 写道:
> The NTFS driver was converting byte offsets to bio sector numbers using
> the filesystem block size (sb->s_blocksize_bits).  However, block layer
> sector numbers - bio->bi_iter.bi_sector and sector_t - are always in
> units of 512 bytes, independent of the filesystem block size.
> 
> On a device with a 4096 byte logical block size (4Kn), sb_min_blocksize()
> lifts s_blocksize to 4096, so every MFT writeback bio built by
> ntfs_sync_mft_mirror(), write_mft_record_nolock() and
> ntfs_mft_writepages() ended up with a sector number eight times too
> small.  The resulting LBA is not only wrong, it is usually not a
> multiple of the logical block size either, and some controllers reject
> the write outright, e.g.:
> 
>    ntfs: (device nvme0n1p3): ntfs_sync_mft_mirror(): I/O error while
>    writing mft mirror record 0x3!
>    ntfs: (device nvme0n1p3): write_mft_record_nolock(): I/O error while
>    writing mft record 0x3!  Marking base inode as bad.
> 
> Reproduced on an Apple SSD AP0256J (4Kn) with a volume using 4096 byte
> sectors, 4096 byte clusters and clusters_per_mft_record = 1; after the
> fix a 13 MB file written to the volume is byte-for-byte identical
> after a umount/mount cycle.
> 
> Fixes: 40796051991d ("ntfs: update in-memory, on-disk structures and headers")
> Signed-off-by: Zhu Tianhao <zhutianhao75@hotmail.com>
> ---
>   fs/ntfs/ntfs.h | 15 ++++++++++++---
>   1 file changed, 12 insertions(+), 3 deletions(-)
> 
> diff --git a/fs/ntfs/ntfs.h b/fs/ntfs/ntfs.h
> index df5a75d506f6..b66ce62f67c9 100644
> --- a/fs/ntfs/ntfs.h
> +++ b/fs/ntfs/ntfs.h
> @@ -20,6 +20,7 @@
>   #include <linux/smp.h>
>   #include <linux/pagemap.h>
>   #include <linux/uidgid.h>
> +#include <linux/blkdev.h>      /* For SECTOR_SHIFT. */
>   
>   #include "volume.h"
>   #include "layout.h"
> @@ -71,7 +72,11 @@
>   #define NTFS_CLU_TO_POFS(vol, clu) (((u64)(clu) << (vol)->cluster_size_bits) & \
>   				    ~PAGE_MASK)
>   
> -#define NTFS_B_TO_SECTOR(vol, b) ((b) >> ((vol)->sb)->s_blocksize_bits)
> +/*
> + * bio->bi_iter.bi_sector and sector_t are always in 512-byte sectors, even
> + * when the filesystem block size is larger (for example 4KiB).
> + */
> +#define NTFS_B_TO_SECTOR(vol, b) ntfs_bytes_to_sector(vol, b)
>   
>   enum {
>   	NTFS_BLOCK_SIZE		= 512,
> @@ -154,11 +159,15 @@ static inline u64 ntfs_cluster_to_poff(const struct ntfs_volume *vol,
>   	return (clu << vol->cluster_size_bits) & ~PAGE_MASK;
>   }
>   
> -/* Convert byte offset to sector (block) number. */
> +/* Convert byte offset to a block layer sector number.  Those are always in
> + * units of 512 bytes, never in units of sb->s_blocksize, which is raised to
> + * the device logical block size on 4Kn devices.
> + */
>   static inline sector_t ntfs_bytes_to_sector(const struct ntfs_volume *vol,
>   		u64 bytes)
>   {
> -	return bytes >> vol->sb->s_blocksize_bits;
> +	(void)vol;
> +	return bytes >> SECTOR_SHIFT;
>   }
>   
>   /* Global variables. */

Hi Tianhao,
    Thank you for submitting this patch.
    This issue has already been fixed by commit 6faa235a649e ("ntfs: 
compute bi_sector in 512-byte units"), which is included in the current 
ntfs-next branch. The existing fix covers the same MFT write paths, so 
this patch does not need to be applied again.
    Thank you for your work.

Thanks,
Baolin.

回复: [PATCH] ntfs: fix sector conversion on 4Kn devices

[朱 天浩]

Hi Baolin,

Thanks for your reply. Commit 6faa235a649e solves the same problem I encountered.

I did test it on MacBookPro14,1: an Apple SSD AP0256J (4Kn), with a volume
using 4096 byte sectors and clusters.  Before the fix, every MFT writeback
fails with -EIO; with 6faa235a649e applied on top of Deepin's customized
6.18.38-amd64-desktop-rolling kernel (where this module is backported),
writing a 13 MB file succeeds and the file is byte-for-byte identical after a
umount/mount cycle.  So for that commit, feel free to add:

    Tested-by: Zhu Tianhao <zhutianhao75@hotmail.com>

Thanks,
Tianhao

________________________________________
发件人: liubaolin <liubaolin12138@163.com>
发送时间: 2026年9月4日 9:32
收件人: 朱 天浩; linkinjeon@kernel.org; hyc.lee@gmail.com
抄送: ntfs@lists.linux.dev; linux-kernel@vger.kernel.org
主题: Re: [PATCH] ntfs: fix sector conversion on 4Kn devices



在 2026/9/3 20:44, 朱 天浩 写道:
> The NTFS driver was converting byte offsets to bio sector numbers using
> the filesystem block size (sb->s_blocksize_bits).  However, block layer
> sector numbers - bio->bi_iter.bi_sector and sector_t - are always in
> units of 512 bytes, independent of the filesystem block size.
>
> On a device with a 4096 byte logical block size (4Kn), sb_min_blocksize()
> lifts s_blocksize to 4096, so every MFT writeback bio built by
> ntfs_sync_mft_mirror(), write_mft_record_nolock() and
> ntfs_mft_writepages() ended up with a sector number eight times too
> small.  The resulting LBA is not only wrong, it is usually not a
> multiple of the logical block size either, and some controllers reject
> the write outright, e.g.:
>
>    ntfs: (device nvme0n1p3): ntfs_sync_mft_mirror(): I/O error while
>    writing mft mirror record 0x3!
>    ntfs: (device nvme0n1p3): write_mft_record_nolock(): I/O error while
>    writing mft record 0x3!  Marking base inode as bad.
>
> Reproduced on an Apple SSD AP0256J (4Kn) with a volume using 4096 byte
> sectors, 4096 byte clusters and clusters_per_mft_record = 1; after the
> fix a 13 MB file written to the volume is byte-for-byte identical
> after a umount/mount cycle.
>
> Fixes: 40796051991d ("ntfs: update in-memory, on-disk structures and headers")
> Signed-off-by: Zhu Tianhao <zhutianhao75@hotmail.com>
> ---
>   fs/ntfs/ntfs.h | 15 ++++++++++++---
>   1 file changed, 12 insertions(+), 3 deletions(-)
>
> diff --git a/fs/ntfs/ntfs.h b/fs/ntfs/ntfs.h
> index df5a75d506f6..b66ce62f67c9 100644
> --- a/fs/ntfs/ntfs.h
> +++ b/fs/ntfs/ntfs.h
> @@ -20,6 +20,7 @@
>   #include <linux/smp.h>
>   #include <linux/pagemap.h>
>   #include <linux/uidgid.h>
> +#include <linux/blkdev.h>      /* For SECTOR_SHIFT. */
>
>   #include "volume.h"
>   #include "layout.h"
> @@ -71,7 +72,11 @@
>   #define NTFS_CLU_TO_POFS(vol, clu) (((u64)(clu) << (vol)->cluster_size_bits) & \
>                                   ~PAGE_MASK)
>
> -#define NTFS_B_TO_SECTOR(vol, b) ((b) >> ((vol)->sb)->s_blocksize_bits)
> +/*
> + * bio->bi_iter.bi_sector and sector_t are always in 512-byte sectors, even
> + * when the filesystem block size is larger (for example 4KiB).
> + */
> +#define NTFS_B_TO_SECTOR(vol, b) ntfs_bytes_to_sector(vol, b)
>
>   enum {
>       NTFS_BLOCK_SIZE         = 512,
> @@ -154,11 +159,15 @@ static inline u64 ntfs_cluster_to_poff(const struct ntfs_volume *vol,
>       return (clu << vol->cluster_size_bits) & ~PAGE_MASK;
>   }
>
> -/* Convert byte offset to sector (block) number. */
> +/* Convert byte offset to a block layer sector number.  Those are always in
> + * units of 512 bytes, never in units of sb->s_blocksize, which is raised to
> + * the device logical block size on 4Kn devices.
> + */
>   static inline sector_t ntfs_bytes_to_sector(const struct ntfs_volume *vol,
>               u64 bytes)
>   {
> -     return bytes >> vol->sb->s_blocksize_bits;
> +     (void)vol;
> +     return bytes >> SECTOR_SHIFT;
>   }
>
>   /* Global variables. */

Hi Tianhao,
    Thank you for submitting this patch.
    This issue has already been fixed by commit 6faa235a649e ("ntfs:
compute bi_sector in 512-byte units"), which is included in the current
ntfs-next branch. The existing fix covers the same MFT write paths, so
this patch does not need to be applied again.
    Thank you for your work.

Thanks,
Baolin.